Repository navigation
scientific: establish longitudinal CWC recovery acceptance evidence #501
Description
Problem
longitudinal_cwc_v1 composes the protected-main Enders–Tofighi CWC owner into an Analysis Run, but the profile-level evidence is still dominated by noiseless/deterministic fixtures. That is not sufficient scientific acceptance for a commercial longitudinal psychometrics path.
Acceptance evidence required
Keep reusable CWC arithmetic in psychometric_core; do not build a second estimator in the acceptance harness. Add a realistic recovery design that exercises the same profile/admission boundary and records, at minimum:
- true within, between, and contextual slopes across repeated generated populations;
- RMSE and bias for all three quantities with Monte Carlo uncertainty;
- attempted / recovered / failed replicate denominators and convergence or refusal rates;
- variation in cluster count, cluster-size imbalance, predictor separation, outcome noise, and between/within signal ratio;
- unequal follow-up or time-varying availability patterns where applicable without collapsing cluster structure;
- leakage-safe rolling-origin or historical-cutoff evaluation when evidence availability varies over time;
- reproducible seed/config identity and exact replay evidence.
Synthetic data may support unit/regression contracts, but product scientific acceptance must report the true-parameter recovery denominators and uncertainty rather than treating a single fixture as evidence of general recovery.
Boundaries
- Do not use an LLM judge for numerical/scientific acceptance.
- Do not weaken the causal-refusal boundary: CWC slopes remain associational without an identification design.
- Do not reinterpret cluster membership or repeated observations as independent atomic evidence.
- Do not move generalized/static psychometric arithmetic out of its canonical owner.
Traceability
Owner PR: #372 (longitudinal_cwc_v1). The PR may repair temporal/provenance/artifact integrity independently, but must not be described as scientifically accepted or release-ready until this issue or equivalent checked-in evidence is satisfied.
Activity
seonghobae commentedon Sep 18, 2026
#592가
longitudinal_cwc_v1의 scientific-acceptance 선행 provenance 결함으로 확인되었습니다. protected-mainexecute_analysis_run은 opaque evidence identity 중복을 fail closed하지만 #372의LongitudinalClusterScore는 identity를 떨어뜨려 같은 observation replay가 CWC row weighting/cluster mean/within·between slope/contextual effect/row_count를 바꿀 수 있었습니다.현재 ordinary-forward lineage는 RED
e66a90f3c6be7c04ecc9a310baf955b16dbd6f76→ production repair138be1fb8ad3ae47a18d7e05645d2ca2830cd341→ fixture migration66cffe0c88f2c8e54881f6018a776c62daaf788c→ exact head418222f232dfc1eb8d6fa27d841e2302148ccf55입니다. 새 계약은 explicitevidence_id, duplicate refusal before cutoff filtering, post-cutoff duplicate refusal, equal numeric/distinct identity positive control을 요구하며 tuple-based dedupe는 금지합니다.이 수리는 #501의 RMSE/bias/MC uncertainty/recovery denominator/cluster-size·noise·availability/rolling-origin acceptance를 대체하지 않습니다. 오히려 그 Monte Carlo/realistic recovery evidence가 동일 observation replay로 오염되지 않도록 입력 population identity를 먼저 고정하는 prerequisite입니다. Exact-head workflows와 focused review는 아직 settling 중이므로 GREEN/acceptance로 승격하지 않습니다.
seonghobae commentedon Sep 18, 2026
Correction/currentization for #592: the evidence-identity finding remains valid, but the first repair over-constrained future-unavailable duplicates. Fresh comparison with the surviving Analysis Run vehicle #416 showed the canonical leakage-safe order is availability admission before duplicate/domain admission.
Current ordinary-forward child lineage is now supplemental RED
bb8cb21ac40fc9f1a2d86612e8cbaaac5ccecb67→ causal repair6b07a4eafa3c009e36299aeca6e87a5b45201d56→ doctoring/current exact head63ac964e9f9c9e94f55a7b881e1ead1ec04a2495.Accordingly the recovery population for #501 must satisfy both: (1) duplicate identity among evidence actually available at the cutoff fails closed so one source observation cannot gain extra CWC weight, and (2) a later-unavailable row, even if it reuses a visible identity, must not perturb an earlier historical replay. This keeps the Monte Carlo/rolling-origin population leakage-safe. #592 remains an integrity prerequisite, not a substitute for #501's RMSE/bias/MC uncertainty/recovery-denominator acceptance.
seonghobae commentedon Sep 18, 2026
#593 adds a second leakage-safe prerequisite for this recovery acceptance lane.
longitudinal_cwc_v1previously excluded post-cutoff rows from numerical fitting but still emittedexcluded_after_cutoff_count, so later corpus additions changed the earlier artifact digest/terminal result. That is incompatible with rolling-origin historical invariance.Current child lineage: RED
35a8f082b87c0f4e9fdd6e41f6f5fee4cd3f601b→ repair28509b74a7949fcfc8b40f479b7cbabdc76b89d1→ fixture migrationsf1230ef23c51865aee6143d3cc7b1e4cb0db3240/4befe868eaf0d2384358c8fffc59432c96f1874b→ ADR98daaa73422f65fd8be9dfdb9e0cd0df50192617→ current doctoring/headac894bdc49ad622ecbe0c105cfb775299c2e6941.Production historical artifacts now intend to expose only cutoff-visible
row_count/cluster_count; future-only attempted/excluded denominators belong in the retrospective #501 acceptance harness, where they can be reported explicitly without contaminating the historical product result. This does not reduce #501's requirement for attempted/recovered/failed denominators, RMSE/bias/MC uncertainty, design variation, and rolling-origin recovery.seonghobae commentedon Sep 18, 2026
#596 is now a numerical prerequisite for this acceptance lane. The CWC owner currently forms each cluster predictor/outcome mean with ordered naive
f64accumulation, so the same six cutoff-visible evidence identities can produce a different between/contextual decomposition when only row enumeration changes. Public consumer REDTEPP#372@2a05f4d38f5136d32ab0170249dd3c47cf95949eaddslongitudinal_cwc_permutation_contract.rsand requires artifact + terminal result bit identity for[1e16, 1, -1e16]versus[1e16, -1e16, 1]ordering of the same cluster-1 evidence. The RED contains no TEPP-local summation workaround.Canonical numerical ownership remains fast-mlsirm.
fast-mlsirm#1816has been ordinary-forward reconciled to protectedmain@a712995b1c22230bc7fcc7f693ae4ad88cb363f4at exact7b853c2e8766c0dd58c3b01064c0dbcad1500493(ahead 19 / behind 0), carrying only its six numerical-owner deltas, but it is still mutable Draft. Therefore #501 recovery acceptance must not treat current CWC permutation sensitivity as sampling noise or absorb it with tolerances; the realistic recovery harness is meaningful only after the releasedfast_mlsirm.binary64_mean@1.0.0contract is consumed for cluster means. #501 still separately requires RMSE/bias/MC uncertainty, recovery denominators, design variation, unequal follow-up/time-varying availability, rolling-origin leakage safety, and replay evidence.seonghobae commentedon Sep 19, 2026
#604 adds a prerequisite to this recovery-acceptance lane: the current
longitudinal_cwc_v1row type carries only onecluster_key, while protected-mainmembership_corepermits time-varying weighted simultaneous/non-nested memberships. Recovery evidence must not sample a multiple-membership/cross-classified truth and then silently flatten it to one group before the Analysis boundary.Keep #501 open and preserve its existing denominators/RMSE/bias/coverage/convergence requirements. Before treating a CWC recovery study as product acceptance, the supported design must be explicit: either (a) authoritative membership state proves the v1 population is genuinely single-membership/nested at the relevant event time, with #604 fail-closed admission for unsupported structures, or (b) a separately versioned cross-classified/multiple-membership estimand is implemented with its own weighting rule and true-parameter recovery. Duplicating rows to represent membership weights is not acceptable because it changes evidence weight and conflicts with #592 anti-pseudoreplication.